paischer101/hypersonic-scramjets

Scramjet Hypersonic Flow Physics Emulator Dataset

A dataset of 6,877 steady-state hypersonic flow simulations of parametrically varied scramjet geometries, produced with JAX-Fluids (https://github.com/tumaer/JAXFLUIDS) as part of a fully GPU-based CFD workflow intended for training and evaluating physics emulators. The current version of the dataset contains the irregular-grid data used for training the AB-UPT emulator (https://arxiv.org/abs/2502.09692) from the paper. We will update the dataset with the regular-grid version used for training ViT and Flow Matching variantes soon along with code to load and pre-train/fine-tune the different emulators.

Each sample pairs a set of geometry + inflow design parameters (train.csv) with the corresponding volume and surface flow fields (run_<i>.zip).

• Paper: A fully GPU-based workflow for building physics emulators of hypersonic flows
• Samples: 6,877 (run_1 … run_6877)

Repository layout

train.csv           # design parameters, one row per sample (6,877 rows + header)
run_1.zip           # flow fields for sample 1
run_2.zip
...
run_6877.zip

Every run_<i>.zip extracts to a directory run_<i>/ containing PyTorch tensor files (torch.save, load with torch.load). All field tensors are float32.

train.csv — design parameters

One row per sample, 18 columns. Row order is 1:1 with the run index: the first data row corresponds to run_1, the second to run_2, …, the last to run_6877.

column	meaning
overall_length	overall vehicle length
flow_path_height_fraction	flow-path height as a fraction of length
cowl_intake_length_fraction	cowl/intake length fraction
airframe_isolator_angle_deg, airframe_combustor_angle_deg	airframe angles (deg)
airframe_length_frac_intake / _isolator / _combustor / _nozzle	airframe section length fractions
intake_ramp_{1,2,3}_angle_deg	intake ramp angles (deg)
intake_ramp_{1,2,3}_length_fraction	intake ramp length fractions
mach_number	freestream Mach number
total_pressure_pa	inflow total pressure (Pa)
total_temperature_k	inflow total temperature (K)

run_<i>.zip — flow fields

Point counts vary per sample (adaptive mesh). Shapes below are for run_1 (N_vol = 1,183,280 volume points, N_surf = 17,801 surface points).

Volume fields (one value per volume point):

file	shape	description
volume_point_positions.pt	(N_vol, 3)	point coordinates (x, y, z)
volume_point_velocity.pt	(N_vol, 2)	velocity components (planar)
volume_point_density.pt	(N_vol,)	density
volume_point_pressure.pt	(N_vol,)	static pressure
volume_point_temperature.pt	(N_vol,)	temperature
volume_point_mach_number.pt	(N_vol,)	Mach number
volume_point_enthalpy.pt	(N_vol,)	enthalpy
volume_point_kinetic_energy.pt	(N_vol,)	kinetic energy
volume_point_total_energy.pt	(N_vol,)	total energy
volume_point_total_pressure.pt	(N_vol,)	total pressure

Surface fields (one value per surface point):

file	shape	description
surface_point_positions.pt	(N_surf, 3)	surface field-point coordinates
surface_positions_geom.pt	(N_surf, 3)	geometry surface coordinates
surface_point_velocity.pt	(N_surf, 2)	velocity components (planar)
surface_point_density.pt	(N_surf,)	density
surface_point_pressure.pt	(N_surf,)	static pressure
surface_point_temperature.pt	(N_surf,)	temperature
surface_point_mach_number.pt	(N_surf,)	Mach number
surface_point_enthalpy.pt	(N_surf,)	enthalpy
surface_point_kinetic_energy.pt	(N_surf,)	kinetic energy
surface_point_total_pressure.pt	(N_surf,)	total pressure

design_parameters.pt — a dict of the 16 geometry/Mach scalars for that sample (a subset of the train.csv columns; train.csv is the authoritative table and additionally includes total_pressure_pa and total_temperature_k).

Usage

import io, zipfile, torch
import pandas as pd
from huggingface_hub import hf_hub_download

REPO = "<user>/scramjet-7000"  # <-- your repo id

# design parameters for every sample (row i -> run_{i+1})
params = pd.read_csv(hf_hub_download(REPO, "train.csv", repo_type="dataset"))

# load the flow fields for one sample
i = 1
zip_path = hf_hub_download(REPO, f"run_{i}.zip", repo_type="dataset")
with zipfile.ZipFile(zip_path) as zf:
    def load(name):
        with zf.open(f"run_{i}/{name}") as f:
            return torch.load(io.BytesIO(f.read()), weights_only=False)

    pos = load("volume_point_positions.pt")   # (N_vol, 3)
    rho = load("volume_point_density.pt")      # (N_vol,)
    design = params.iloc[i - 1]                # matching design parameters

Citation

If you use this dataset, please cite:

@misc{Paischer2026,
  title={A fully GPU-based workflow for building physics emulators of hypersonic flows},
  author={Paischer, Fabian and Rubini, Dylan and Bezgin, Deniz A. and Buhendwa, Aaron B. and Hauser, David and Sestak, Florian and Brandstetter, Johannes and Kaltenbach, Sebastian and Adams, Nikolaus A.},
  year={2026},
  eprint={2606.13742},
  archivePrefix={arXiv},
  primaryClass={cs.LG}
}